Scanning with a grid of candidates, not one glance
Conceptually, a modern detector doesn't look at an image once and instantly report objects. It processes the image through convolutional layers that build up a spatial feature map, essentially a grid laid over the image where each cell summarizes what's happening in that patch and its surroundings. At each grid cell, or at a set of predefined anchor shapes centered on that cell, the model asks: does this look like it contains part of an object, and if so, what class and what refined box would fit it best.
Convolutional filters are what make this grid of local judgments possible in the first place. Each filter slides across the image looking for a specific local pattern, an edge, a texture, a shape fragment, and stacking many filters across many layers lets the network build from raw pixels up to 'this patch looks like part of a car.' That's the same filter mechanism used in basic image classification, just applied densely across the whole image instead of once for the whole thing.